LLMs aren't actually able to do that though, are they? They are simply incapable of keeping track of consistent behaviors and beliefs. I recognize that for certain prompts an LLM has to do it. But as long as we're using transformers, it'll never actually work.